On using physico-chemical properties of amino acids in string kernels for protein classification via support vector machines

String kernels are popular tools for analyzing protein sequence data and they have been successfully applied to many computational biology problems. The traditional string kernels assume that different substrings are independent. However, substrings can be highly correlated due to their substructure relationship or common physico-chemical properties. This paper proposes two kinds of weighted spectrum kernels: The correlation spectrum kernel and the AA spectrum kernel. We evaluate their performances by predicting glycan-binding proteins of 12 glycans. The results show that the correlation spectrum kernel and the AA spectrum kernel perform significantly better than the spectrum kernel for nearly all the 12 glycans. By comparing the predictive power of AA spectrum kernels constructed by different physico-chemical properties, the authors can also identify the physicochemical properties which contributes the most to the glycan-protein binding. The results indicate that physico-chemical properties of amino acids in proteins play an important role in the mechanism of glycan-protein binding.